We present a configurable pipeline and the associated code that can be used to generate multilingual sets of entities with specified characteristics, such as domain, geographical location and popularity, using data from Wikipedia and Wikidata. These datasets are intended for evaluating the factuality of LLMs’ long-form generation, thereby complementing evaluation based on short-form QA datasets. We present the RiDiC dataset as an example of this approach. RiDiC contains 3,000 entities from three domains – rivers, natural disasters, and car models – spanning different popularity tiers. Each entity is accompanied by its geographical location, English and Chinese names (if available) and relevant English and Chinese Wikipedia content, which is used to evaluate LLMs’ responses. Generations about RiDiC entities were obtained from three LLMs in English and Chinese. These were then evaluated using a third-party factuality checker, which showed that entities from our dataset caused even frontier models to hallucinate. The code, data and generation/evaluation scripts have been released to enable the approach to be extended to new LLMs, languages and domains.

Recent work in Graph-to-Text generation has achieved impressive results, but it still suffers from hallucinations in some cases, despite extensive pretraining stages and various methods for working with graph data. While the commonly used metrics for evaluating the quality of Graph-to-Text models show almost perfect results, it makes it challenging to compare different approaches. This paper demonstrates the challenges of recent Graph-to-Text systems in terms of hallucinations and proposes a simple yet effective approach to using a general LLM, which has shown state-of-the-art results and reduced the number of factual hallucinations. We provide step-by-step instructions on how to develop prompts for language models and a detailed analysis of potential factual errors in the generated text.